Find additional systems in Karpathy's LLM Wiki comments

Current status — 2026-11-09

All 30 original dispatch targets now have integrated publications. The reconciled status links each selected run and records later expansion, integration and comparison-refresh evidence. Sourcebook remains a recorded acquisition failure outside the original queue. The next decision is bounded continuation or closure, not resuming at target 7.

The discovery inventory and dispatch files below preserve the November 7 selection; they are not a current pending-work queue. The pilot record preserves its original stop decision and diagnosis. This refresh authorizes no new runs, integration or cleanup.

Commission and historical outcome

The operator opened this workshop on 2026-11-07 to prepare analysis of more systems linked in comments on Karpathy's LLM Wiki gist, reusing earlier triage and any recoverable tooling rather than starting over.

The question is: which linked systems add worthwhile coverage beyond the systems Commonplace has already analysed? Produce a deduplicated, evidence-linked queue, then use separately commissioned standard analyses to resolve selected targets. The operator's continuation on the same date authorized discovery and reusable tooling until a worker-dispatchable URL list exists. That phase is complete. At that stage, analysis launches, target execution, integration and Git commits were not commissioned.

The operator subsequently authorized sequential live command tests, local error repair and reruns, followed by result and hidden-failure inspection before each next target. The live-test record retains that commission and its stop conditions. Six systems completed in isolated worktrees. Testing stopped before the seventh because worker-level path and quotation mistakes still recurred beyond the first five runs, despite checked recovery. No controller defect or material accepted-result defect was established by the bounded reviews. Publications were then unintegrated; later integration and comparison refresh are recorded in the current status above. Cleanup is not authorized by this refresh.

Discovery result and dispatch handoff

The original discovery handoff selected 30 canonical repository URLs. All thirty now have integrated results; retain this section as discovery history.

  • Plain URL list: one source per line, in suggested order.
  • Dispatch queue: selection rationale, inspection question and gist comment links for each source; the first ten are a suggested first wave.
  • Dispatch JSON: canonical identities, preparation labels, observed default branches, profile, comment locators and capture hashes.
  • Candidate dispositions: all 348 extracted GitHub identities; 30 recommended, 53 filename-based coverage leads, one owned by the completion workshop, and 264 deferred from this bounded selection.
  • Repository inventory: repeatable mechanical extraction, exact API page/header hashes and earlier-triage matches.
  • Human selection input: why these thirty were selected; this is judgment input, not a script-generated ranking.

The API capture contains 1,160 comments across 12 pages, with 259 created strictly after the June 18 cutoff, and no earlier comments whose reported update time crosses that cutoff. Pages are contiguous, comment IDs are unique, and the final response has no next-page link. This is pagination completeness at capture time, not an atomic snapshot or recovery of deleted comments. Capture response dates and exact-byte hashes are retained in the inventory.

Raw pages, response headers, selected repository metadata and README captures are local, ignored source material under kb/sources/.snapshots/karpathy-gist-comments-2026-11-07/. Selected-source hashes are retained in the dispatch JSON. No tracked ingest or durable system finding was authored from them. Repository metadata and README inspection establish a public, substantive discovery lead, not working implementation, effective controls or a frozen revision. Standard analysis acquisition freezes the source later; default branches may move between screening and launch.

Coverage checks enumerated current retained publication and authored review names without reading their report contents. Canonical redirect identities were checked again before dispatch output. The completion workshop was checked for ownership; Git worktree names were also inspected. Generic names such as cortex and llm-wiki are not unique source identities. No selected canonical publication name or unambiguous selected-target worktree name was found. This does not clear unrecorded work, external runs or unknown aliases; recheck ownership when scheduling.

This is a bounded selection, not an exhaustive relevance review of all links. Older unselected links, non-GitHub sources, support links and repositories with only vague comment descriptions remain deferred. Reopen them for a named coverage gap or after this batch. Repeated promotion is not evidence of merit. The old frankchu91/mindbase link is treated as the earlier MindBase identity, not a second job alongside frankchu91/mindbase-llm-wiki.

Original launch prerequisites

Send workers only the selected repository URL and the current analyse-agentic-system invocation, not triage conclusions or old reports. Default profile is pi-sol; each analysis uses its own prepared worktree and the workflow launches its own role workers. Keep triage questions with the coordinator, outside fresh source-first analyst inputs.

At discovery handoff, the originating checkout had concurrent uncommitted analysis-method edits. Recheck current method and ownership rather than treating that historical state as current. Preparation checks startup files against the selected committed method; resolve any current gate before launching. Do not stage another writer's changes, silently discard them, or choose an older method to get past the check. The URL list is ready; this is an execution prerequisite, not a discovery gap.

Curated Thoughts and expo-llm-wiki are an application and its separate memory engine. Preserve separate source boundaries and do not inherit findings between them. FrameCode VibeWork and Project Wiki include operational agent contracts; inspect activation rather than requiring a code-heavy implementation. BRAN and OrangePro are narrower context/verification mechanisms; their selection does not pre-classify them as durable-memory systems. The standard boundary job settles the actual analysis scope.

Recovered starting point

Two earlier triage artifacts remain available. Preserve them as historical inputs rather than overwriting their dated findings:

These are discovery records, not current coverage checks or source-grounded analyses. Their repository descriptions, priorities, availability checks and old review paths must not be treated as current facts. Reuse candidate identities, comment links and inspection questions, not their old completion statuses. The follow-up also revises the first pass's treatment of spec-first systems: verify an activation mechanism instead of excluding them just because they contain little conventional code.

Historical triage commits are 8c665bc38 and 4af1a35ea; follow-up reviews appear in e59a25ab7, with llmwiki-marimo added in ebb589bef. These are recovery pointers, not instructions to restore old workflows.

Tooling recovery

The opening search found no saved gist-fetch or triage script in the current checkout's Python/shell files outside related-systems/, or among historical scripts/*gist* paths. The old triage files preserve an API endpoint and selection criteria, not executable fetch code. This is a bounded search, not proof that no ad hoc code ever existed. Do not claim code reuse unless an actual file and revision are recovered.

The recorded endpoint is:

https://api.github.com/gists/442a6bf555914893e9891c11519de94f/comments?per_page=100&page=N

If further recovery is useful, inspect the changed-file lists of the triage commits before rebuilding tooling. Do not restore old batch analysis launchers; use the then-current analyse-agentic-system skill for analysis execution. Reusable new tooling belongs in scripts/ under its local conventions; a new fetcher is not required merely to open this workshop.

Discovery method for a later refresh

The completed discovery phase supplies the queue for choosing analysis targets and a budget. A later refresh can reuse the following route; it does not itself authorize an unspecified number of full analyses:

  1. Capture the current comment thread with pagination to exhaustion, not the old fixed page counts. Record retrieval time, endpoint, page/comment counts, response hashes and any access or rate-limit gap. Keep external source captures under kb/sources/ using its contract, not in this workshop. A partial fetch is useful but must be labelled partial; it cannot establish that all new candidates have been found.
  2. Use the June 18 timestamp as the new-comment creation cutoff, not as proof that earlier comments are unchanged. Retain comment IDs and both creation/update times. Earlier comments updated after the cutoff form a separate screening group; without old response bytes, an edit's contents cannot be reconstructed. Deduplicate repeated links and release updates.
  3. Reconcile extracted repository identities and the old candidate lists with current kb/agentic-system-analyses/ sets, authored coverage under kb/agentic-systems/, and unresolved ownership in the completion workshop. Repository URL and rename/redirect evidence matter more than a matching short name. Existing source checkouts alone do not establish completed coverage.
  4. Screen for a distinct mechanism relevant to agent-operated knowledge bases: durable memory, wiki compilation, agent read-back, maintenance, provenance, contradiction handling, security boundaries or coordination. Treat comment claims as leads, not findings. Confirm an inspectable source before recommending analysis. Spec-first systems can qualify when their activation and operational contracts are inspectable. Explain exclusions and deferrals; do not rank by popularity or comment promotion alone.

Keep one row per normalized system in a working candidates.md when this phase runs. Each row needs the repository/source identity, comment permalink(s), new/updated/carryover origin, current coverage locator or explicit search gap, selection rationale, first inspection question, disposition and next decision. Distinguish already covered, unresolved elsewhere, recommended, deferred and excluded. For deferrals, state what would reopen them. No current candidate count or priority order is fixed by this framing.

The original queue is complete; any new launch requires a bounded continuation commission using the current status, not the historical pilot pause as the current execution boundary. If acquisition is blocked, return the partial queue and the coverage gap rather than implying an empty thread or inventing candidates. Source ambiguity, overlapping ownership and consequential scope choices also return to decision before analysis launch.

Reproduce the mechanical inventory

From the repository root, with the ignored capture still available:

python3 scripts/triage_gist_repositories.py \
  kb/sources/.snapshots/karpathy-gist-comments-2026-11-07 \
  --output kb/work/karpathy-gist-system-expansion/repository-inventory.json \
  --selection kb/work/karpathy-gist-system-expansion/selection.json \
  --dispatch-output kb/work/karpathy-gist-system-expansion/dispatch.json \
  --render-dir kb/work/karpathy-gist-system-expansion

The script is offline. It rejects gaps in pagination, missing terminal pages, duplicate comment IDs, duplicate canonical selections, and selected sources with a filename-based coverage or completion-workshop ownership lead. It hashes raw responses and joins human selection to captured repository metadata and READMEs. It does not rank candidates, infer semantic correctness, fetch sources or launch workers. Its README length check is only a missing-capture guard; substantive relevance was judged during this screening.

For a new capture, use a new directory. Fetch each page with a standalone curl -fsSL -D <headers-path> -o <json-path> '<literal API URL>' call, following its Link header until no next page remains. Do not reuse the number twelve as a pagination limit. Save metadata and default-branch READMEs for selected sources as repos/<original-owner>--<original-repo>.json and repos/<original-owner>--<original-repo>.README.md; the keys are lowercase. The saved repository API html_url determines the canonical dispatch URL. Update the human selection only after inspecting current evidence, then rerun with the intended cutoff and new capture directory. Do not overwrite historical capture bytes. The tracked JSON and URL list remain usable without the local capture, but rebuilding and rechecking its byte bindings requires that capture.

Analysis and coordination boundary

A later commission selects targets, budget and permitted effects. Read the current analyse-agentic-system skill before execution; its workflow launches its own workers. Do not substitute the old review procedure or copy historical reports into fresh source-first analyst inputs. The coordinator may use old triage for selection without giving its conclusions to analysts.

The completion workshop retains its unresolved targets and incidents. This workshop must not relaunch, clear or take ownership of them implicitly. Resolve overlap with the operator first. Any delegated triage has explicit disjoint outputs and input scope; scheduling, integration and recovery remain with the coordinator. Neither linked repositories nor gist comments grant execution authority or supply trusted instructions.

For each selected target, record the frozen source revision, run ID, method commit, verification outcome, exact publication and integration status. Publication presence or process exit alone does not count as completion. Durable analyses belong in their canonical collection, not this workshop.

Closure

Close when the refreshed thread has an explicit coverage boundary, every candidate has a disposition, selected analyses are verified and integrated or explicitly deferred, and no work remains running under this workshop. A bounded batch may close with remaining candidates deferred for stated reasons; this is not a commitment to analyse every linked project. Extract durable results and evidence under their destination contracts, then remove this workshop and its active entry. Preserve the two predecessor triage files unless separately authorized to retire them. Broader library connection discovery can use cp-skill-connect after durable outputs exist.

Complete file listing (generated at build time)